Vocal separation usually means picking one model and tolerating its artifacts — a hollow lead vocal here, a smeared cymbal there. The quiet insight behind this project is that no single network wins on every track, so it treats separation as a curation problem: it wraps three distinct lineages — Meta's Demucs (v3/v4), the MDX-Net / MDX23C spectrogram models, and tsurumeso's older VR architecture — behind one interface and lets you ensemble and A/B them.
What Sets It Apart
- Ensemble across model families, not just versions: when MDX-Net leaves a vocal residue, a Demucs pass or an averaged ensemble often removes it — you trade compute time for cleaner stems.
- Everything runs locally, so there are no upload limits, no per-track fees, and no audio leaving your machine — practical for full albums or copyrighted material.
- Hardware reach is unusually wide: CUDA, AMD Radeon, Intel Arc, and Apple Silicon (MPS), so the same workflow runs on a gaming PC or an M-series Mac.
- The knobs that actually move quality — segment/window size, denoise, and Rubber Band pitch-shift / time-stretch — are exposed rather than hidden.
Who It's For
Great fit if you separate stems regularly — producers building acapellas or instrumentals, remixers, karaoke creators, or anyone who wants to run several models on the same file and keep the best take. Look elsewhere if you need a one-click web service, a batch API, or real-time separation; this is a hands-on desktop GUI built for quality-first offline work, and the cleanest results often come from running and comparing multiple models.
